Intel Diamond Rapids Xeon: 256 P-Cores, 1.6TB/s Bandwidth, Still No Launch Date
- Harold Fritts
- 313 views
Intel's SC26 outline for Diamond Rapids claims 256 P-cores, 1.6TB/s of memory bandwidth and 1.28GB of last-level cache, but leaves key memory details unstated.

Intel used SC26 to put Diamond Rapids, its next-generation Xeon server architecture, at the front of its high-performance computing pitch. The headline numbers are large: up to 256 performance cores, up to 1.6TB/s of memory bandwidth and up to 1.28GB of last-level cache. Intel hasn't given a launch date, though, and several of the figures that would let a buyer size a system are still missing. The outline came from Jeff McVeigh, vice president and general manager of Data Center Strategic Programs, and it covered the Intel GPU D-series, HPC software, silicon photonics and quantum development alongside the CPU.

What the Diamond Rapids numbers do and don't tell you
Intel describes Diamond Rapids, which it detailed at Hot Chips, as an 18A-P design. The pitch is higher compute density, more memory bandwidth, larger caches and expanded CPU acceleration. Intel Advanced Performance Extensions (APX) are aimed at per-thread performance. The 1.6TB/s figure comes from what Intel calls a unified memory fabric.
Here's the thing: that bandwidth number is hard to plan around yet. Intel hasn't said how many memory channels sit behind it, what memory speed it assumes, or whether 1.6TB/s is a per-socket figure. Those three answers determine whether the number is useful to a capacity planner or just a headline. Until they arrive, treat it as a direction rather than a spec.
For core count, Intel's reference point is AMD. The top AMD EPYC Venice parts that StorageReview has reviewed carry 256 Zen 6c cores per socket. So 256 P-cores puts Diamond Rapids on the same headline count, and the comparison that matters will come down to per-core performance, memory subsystem behavior and power, none of which Intel has yet quantified for this platform.
The architecture also supports Intel AVX 10.2 and enhanced Intel Advanced Matrix Extensions (Intel AMX). Intel calls the chip the only x86 server CPU with built-in matrix compute for HPC and AI workloads. That's a claim about a feature set, and it's worth checking against the instruction support your code actually relies on once software is available to test.
McVeigh framed the roadmap around mixed systems, where CPUs orchestrate workflows and run parallel communication libraries while accelerators take on the AI work. He wrote that the future of HPC is "open, heterogeneous and balanced." That positioning matters because it tells you where Intel expects the CPU to sit in a rack: less as the sole engine, more as the coordinator around accelerators.
GPU D-series: memory capacity as the pitch
The Intel GPU D-series combines AI acceleration with double-precision (FP64) compute in an air-cooled PCIe card. Datatype support runs from FP4 and MXFP4 up through FP64. The Intel-branded GPU D270 carries 160GB of LPDDR5X with 1.5TB/s of memory bandwidth, and partner-designed configurations scale to 480GB.
Intel's argument for research sites rests on memory capacity. Keeping larger working datasets on the card cuts down on partitioning and host-to-accelerator transfers, which is often where scientific workflows lose time. Intel also pitches the card for always-on AI agents that generate code and launch compute jobs, a use case that leans on large resident memory more than on peak throughput.

At its SC26 booth, Intel is running two live demos of the GPU D270, the card formerly code-named Crescent Island. The first is a seismic imaging mini-application from Veridien, ported from HIP and CUDA to SYCL in less than a month. Intel says 90% of its kernels ran well with no Intel-specific optimization. That figure is worth noting for anyone with a CUDA or HIP codebase, because the porting effort is the real cost of moving, not the hardware price. Still, one application's porting experience is a single data point.
The second demo splits an inference pipeline across the two kinds of hardware. D270 cards handle prefill, SambaNova SN50 units handle decode, and Xeon 6 and Xeon 6+ processors orchestrate the workflow. It's a concrete example of the heterogeneous model McVeigh described, and it shows Intel's CPUs staying in the control path even when accelerators do the heavy lifting.
Benchmarks Intel chose to publish
Intel compared the D270 with NVIDIA's RTX PRO 6000 Blackwell Server Edition. On FP64 financial workloads, Intel claims up to 5.9x the throughput on European Monte Carlo, 5.6x on European Black-Scholes, 4.6x on binomial pricing and 1.92x on American Black-Scholes. The D270 is rated at 350W and the NVIDIA card at 600W. For inference, Intel claims up to 29% lower latency on Drex v2.0 and 1.12x the output token throughput on Qwen3-30B-A3B at FP8.
These are Intel's numbers, measured on pre-production silicon, and they haven't been independently verified. The financial test configuration lists eight NVIDIA cards against a single D270, and it doesn't say whether results were normalized per card. The NVIDIA inference runs were done on an AWS virtual machine. If you're building a purchase case, ask for the full test configuration and rerun the workloads that matter to you, rather than lifting the ratios into a slide.
Software and optics
On software, Intel says it will keep investing in Intel MPI, Fortran, compiler optimizations and oneAPI, so existing parallel codebases keep performing on mixed CPU and GPU systems. It's also working on upstream integration with AI frameworks, inference runtimes and deployment environments. The stated goal is for standard models and tools to run on new hardware such as the D-series from the first day it's deployed. For HPC shops with years of MPI and Fortran code, that continuity matters more than any single benchmark.
Intel also pointed to silicon photonics, which uses light to move data over longer distances at what the company describes as a fraction of the energy of copper links. No figures or product timeline came with that claim, so it's a direction to watch rather than something to plan against.
Quantum and what to watch next
On quantum, Intel restated its case for the planned purchase of Classiq. The deal would combine Classiq's automated quantum algorithm synthesis platform with Intel's silicon spin-qubit technology and chip manufacturing. Intel has said it expects the deal to close in the first half of 2027, pending final approvals.
My read is that the near-term decisions for most data centre teams sit with Diamond Rapids' memory subsystem and the D-series' memory capacity, not with quantum. Before Intel publishes a launch date, the questions to put to the vendor are concrete: channel count and memory speed behind the 1.6TB/s figure, whether that number is per socket, and how the 256-core part's power envelope compares with what you run now. Until those answers exist, the sensible move is to inventory which HPC codes depend on MPI, Fortran and AVX-class instructions, so you can test them the day hardware reaches your lab.